A year ago, DeepMind’s AlphaFold AI changed the shape of science — but there is more work to do

OpenAI’s ChatGPT may have captured the AI zeitgeist last fall, but it was DeepMind’s AlphaFold AI that shook the science world last summer.

A year ago, on July 28, 2022, the Alphabet-owned company announced that AlphaFold had predicted the structures for nearly all proteins known to science and dramatically increasing the potential to understand biology — and, in turn, accelerate drug discovery and cure diseases. That built on its groundbreaking work from a year earlier, when DeepMind open-sourced the AlphaFold system that had mapped 98.5 percent of the proteins used in the human body.

Today, DeepMind (now Google DeepMind) says the AlphaFold Protein Structure Database has been used by over 1.2 million researchers in over 190 countries, and that adoption rates of AlphaFold are growing fast in all domains.

A few weeks ago, DeepMind (now Google DeepMind) CEO Demis Hassabis told The Verge that while AI chatbots have gone viral, he believes it is AlphaFold that has “had the most unequivocally biggest beneficial effects so far in AI on the world.” Nearly every biologist in the world has used it, he pointed out, while Big Pharma companies are using it to advance their drug discovery programs.

“I’ve had multiple, dozens, of Nobel Prize-winner-level biologists and chemists talk to me about how they’re using AlphaFold,” he said, while admitting that “the average person in the street doesn’t know what proteins are…whereas obviously, for a chatbot, everyone can understand, this is incredible.”

DeepMind continues to invest in AlphaFold

Of course, in an era when top AI companies are dealing with potential regulation, a rising tide of lawsuits and criticism about model risks, it helps to have a big win with AI that offers unequivocal benefits to humanity. According to DeepMind, AlphaFold has already been used to discover new disease threats in Madacascar; develop a more effective malaria vaccine; develop new drugs to treat cancer; and tackle antibiotic resistance.

But the AlphaFold team isn’t resting on its laurels: One of AlphaFold’s researchers, Kathryn Tunyasuvunakool, told VentureBeat in an interview that “there are a lot of problems in proteins that are not fully solved,” and that it would be “wonderful” to see more real-world applications for AlphaFold over the next 10-20 years.

“I just want to see AI continuing to make a positive impact on problems in biology,” she said. “It’s such a complicated field with such messy data, and it really feels like the sort of thing where we need computers to help us unpick how this all fits together.”

DeepMind is no longer alone in its shape-shifting science prediction efforts: In November 2022, Meta used an AI language model to predict the structures of more than 600 million proteins of viruses, bacteria and other microbes. And it was able to make those predictions in just two weeks.

However, Hassabis said on a recent podcast with Ezra Klein that “advancing science and medicine is always going to be at the heart of what we do and our overall mission…that involves us continuing to invest and work on scientific problems like AlphaFold.”

DeepMind’s AlphaFold solved the ‘protein-folding challenge’

DeepMind had actually first solved what was a half-century-long biology conundrum — known as the “protein-folding challenge” — in November 2020, when it first released AlphaFold.

Proteins, which support nearly all of life’s functions, are complex molecules made up of chains of amino acids, each with its own unique 3D structure. Figuring out how proteins fold into their unique crumpled shapes had been a persistent problem, but AlphaFold offered a new method to accurately predict those structures. The system was trained on the amino acid structures of 100,000-150,000 proteins.

“It’s by far the most complicated system we ever worked on,” Hassabis told Klein. “And it took five years of work and many difficult wrong turns.”

Tunyasuvunakool said that she was one of the “more pessimistic” people on the AlphaFold team. “I was not at all confident that this is a problem that we will be able to solve — I never really imagined we would get to this sort of impactful level of accuracy,” she said. “It was only later that I started to think, if we actually solve this, this is going to be quite a big deal.”

The biggest problem, she said, was the sheer magnitude of different options for how a protein can fold if it wants to go from a linear sequence of amino acids to a complex 3D structure. “There are just billions and billions of combinations for how that structure could look.”

In July 2022, DeepMind announced that AlphaFold had predicted more than 200 million protein structures, which was nearly all of those catalogued on a globally-recognized repository of protein research.

According to DeepMind, a single protein structure can take the whole length of a PhD and cost an average of $100,000 to determine experimentally. By predicting the structures of over 200 million proteins, AlphaFold “potentially saved the equivalent of up to 1 billion years of research and trillions of dollars.”

There are plenty of protein problems left to solve

Tunyasuvunakool emphasized that while AlphaFold solved one big challenge, there are still plenty of “holy grail” problems in the world of proteins that are not fully solved.

“A better understanding of protein physics would be a big one,” she said, explaining that AlphaFold mainly predicts static protein structures, but a lot of proteins perform their function by changing their shape over time.

“So if you think about something like a channel that decides whether to let things in and out of the cell, those tend to come into two different shapes — and for certain applications, you really care about having this structure versus this one, or knowing about how much time they spend in each of those states,” she said. Understanding that distribution is important for areas like medicine and drug development, she explained: “Having a model that is more aware of protein physics, that was able to predict the multiple states that a protein moves through would be really helpful.”

Overall, she said, the biggest excitement is around seeing the level of uptake of AlphaFold as a tool across the field of biology.

“I think it’s pretty unusual for computational biology tools to make this much of a widespread impact.” she said. “At this stage, the paper has had over 10,000 citations — I think I can comfortably say it’s going to be the biggest thing I ever work on.”

But DeepMind likely has larger ambitions in the space: In 2021, Hassabis launched the biotech startup Isomorphic Labs for drug research, which is reportedly getting “closer to securing its first commercial deal” and is “building on the AlphaFold breakthrough as DeepMind’s sister company.”